fix: use the GRPO trainer for evaluation

Author: baixiac

app/api/utils.py

@@ -27,7 +27,13 @@
 from fastapi_users.jwt import decode_jwt
 from app.config import Settings
 from app.domain import TagsGenerative
-from app.exception import StartTrainingException, AnnotationException, ConfigurationException, ClientException
+from app.exception import (
+    StartTrainingException,
+    AnnotationException,
+    ConfigurationException,
+    ClientException,
+    ExtraDependencyRequiredException,
+)
 
 logger = logging.getLogger("cms")
 
@@ -118,6 +124,24 @@ async def configuration_exception_handler(_: Request, exception: ConfigurationEx
         logger.exception(exception)
         return JSONResponse(status_code=HTTP_500_INTERNAL_SERVER_ERROR, content={"message": str(exception)})
 
+    @app.exception_handler(ExtraDependencyRequiredException)
+    async def extra_dependency_exception_handler(
+        _: Request,
+        exception: ExtraDependencyRequiredException
+    ) -> JSONResponse:
+        """
+        Handles extra dependency required exceptions.
+
+        Args:
+            _ (Request): The request object.
+            exception (ExtraDependencyRequiredException): The extra dependency required exception.
+
+        Returns:
+            JSONResponse: A JSON response with a 500 status code and an error message.
+        """
+        logger.exception(exception)
+        return JSONResponse(status_code=HTTP_500_INTERNAL_SERVER_ERROR, content={"message": str(exception)})
+
     @app.exception_handler(ClientException)
     async def client_exception_handler(_: Request, exception: ClientException) -> JSONResponse:
         """
@@ -299,8 +323,8 @@ async def init_vllm_engine(app: FastAPI,
         )
         from vllm import SamplingParams, TokensPrompt
     except ImportError:
-        # Raise a custom exception if vLLM is not installed
-        raise ConfigurationException("Cannot import the vLLM engine. Please install it with `pip install vllm`.")
+        logger.error("Cannot import the vLLM engine. Please install it with `pip install cms[vllm]`.")
+        raise ExtraDependencyRequiredException("Cannot import the vLLM engine. Please install it with `pip install cms[vllm]`.")
 
     parser = FlexibleArgumentParser()
     parser = make_arg_parser(parser)

app/exception.py

@@ -32,3 +32,7 @@ class DatasetException(Exception):
 
 class DeviceNotAvailableError(RuntimeError):
     """An exception raised when a specificy device is required but not available."""
+
+
+class ExtraDependencyRequiredException(Exception):
+    """An exception raised when an extra dependency is required but not found."""

app/model_services/huggingface_llm_model.py

@@ -369,8 +369,9 @@ def create_embeddings(
         sum_hidden_states = masked_hidden_states.sum(dim=1)
         num_tokens = attention_mask.sum(dim=1, keepdim=True)
         embeddings = sum_hidden_states / num_tokens
+        l2_normalised = torch.nn.functional.normalize(embeddings, p=2, dim=1)
 
-        results = embeddings.cpu().numpy().tolist()
+        results = l2_normalised.cpu().numpy().tolist()
         return results[0] if isinstance(text, str) else results
 
     def train_supervised(